GRAPH CONVOLUTIONAL NETWORKS & ADVERSARIAL TRAINING FOR JOINT EXTRACTION OF ENTITY AND RELATION

Entity recognition and relation extraction are the core tasks in information extraction. Currently, supervised deep learning extraction methods are mainly divided into two categories: pipeline and joint entity-relation extraction. The pipeline method has problem of exposure bias, information redunda...

Full description

Saved in:
Bibliographic Details
Published inScientific Bulletin. Series C, Electrical Engineering and Computer Science no. 3; p. 213
Main Authors Qu, Xiaolong, Zhang, Yang, Tian, Ziwei, Li, Yuxun, Li, Dongmei, Zhang, Xiaoping
Format Journal Article
LanguageEnglish
Published Bucharest University Polytechnica of Bucharest 01.01.2023
Subjects
Annotations
Artificial neural networks
Graph neural networks
Information retrieval
Machine learning
Redundancy
Training
Online AccessGet full text

Cover

More Information
Summary:Entity recognition and relation extraction are the core tasks in information extraction. Currently, supervised deep learning extraction methods are mainly divided into two categories: pipeline and joint entity-relation extraction. The pipeline method has problem of exposure bias, information redundancy, error accumulation and interaction missing. To solve the problems, researchers proposed joint entity-relation extraction method. However, the joint entity-relation extraction method based on sequence annotation does not effectively process entity overlapping, and relation overlapping. Therefore, we propose a joint extraction model GcnJere based on graph convolutional neural network to solve existing problems in the pipeline method and further improve the processing effect of entity overlapping and relation overlapping. Furthermore, we combine the advantages of adversarial training and propose GcnJereAT to improve the generalization ability and robustness of GcnJere. Finally, the performance of the proposed two models is verified in the public benchmark dataset. The experimental results indicate that the computational performance of the two models is superior to the comparison models.
ISSN:2286-3540